Human Trajectory Prediction with Stabilization Control

Human motion trajectories hold significant reference value for vehicle path planning. Predicting realistic human motion trajectories amidst uncertainty is of practical importance for vehicle and robot motion planning. Recently, the emerging diffusion models have demonstrated substantial generative capabilities and have been widely applied in image generation and video synthesis. Given their potential to generate predictions that align with human expectations, Denoising Diffusion Probabilistic Models (DDPM) are increasingly being leveraged in the field of trajectory prediction. In this paper, we employ DDPM for trajectory prediction, yielding rational and realistic predictions. To address the issue of slow inference speed, we employ leapfrog technique. Furthermore, we design a model based on the Transformer architecture to capture dependency information. Since non-stabilization data can lead to inaccuracies in Transformer predictions, we propose a stabilization module to mitigate this issue. The stabilization module reduces the non-stabilization of the original sequence, resulting in smoother predictions. However, such smoothing can impair the accuracy of predictions for sudden movements. To counteract this, we introduce a non-stabilization module at the end to restore diversity and ensure the complexity of human trajectories. Our method achieves outstanding results on the NBA dataset, demonstrating the superiority of our model.